Problems with a purpose
Implement the pieces you actually use: broadcasting, gradient descent, normalization, attention, and more.
From your first tensor to a transformer from scratch. Build real ML intuition, one working implementation at a time.
Turn the equation into something that runs.
def attention(q, k, v):
# All you need is a few good primitives.
d = q.shape[-1]
scores = q @ k.transpose(-2, -1)
scores = scores / math.sqrt(d)
weights = torch.softmax(scores, dim=-1)
return weights @ vReading a paper and implementing it are different skills. This is where you practice the second one.
Implement the pieces you actually use: broadcasting, gradient descent, normalization, attention, and more.
Run sample checks as you work. Submit against the full test suite and see exactly where your next attempt should focus.
Follow a study plan, save a tricky problem, and pick up your code where you left it. Your progress stays with your account.
Sign in with Google to request access. The owner approves each account before practice or AI sessions become available.